Skip to main content
This walks you through the fastest path to a finished result in Cowork: create a worker, hand it a goal in chat, and watch it work. If you want the mental model first, read Core concepts; this page assumes you just want to get moving.

​
Prerequisites

  • A CogniAgent workspace you’re a member of. Workers and their tasks live inside a workspace — every member can see and use the same workers.
  • Nothing to install. Workers run in a hosted sandbox in the cloud, not on your machine.
  • Optional: accounts you want a worker to use (Gmail, Slack, a CRM, and so on). You can connect these during setup or add them later — nothing here is a hard blocker to starting.

​
Build and run your first worker

1

Pick a starting point

Click Hire from the Roster to create a worker. Choose a prebuilt template (a ready-made role with starter instructions, skills, and suggested connections) or pick Blank to build one from scratch. Either way you land in the same setup flow next.
2

Name it and set instructions

Give your worker a name, an avatar, and a one-line role. If you started from a template, its starter instructions are already filled in — read them over and edit anything you’d change. Instructions are the persona and ground rules the worker follows on every task; you can always refine them later. See Anatomy of a worker for what else lives here.
3

Attach skills and connections

Review the starter skills (packaged how-to guides the worker loads when it needs them) and connections (integrations like email, a CRM, or a calendar) the template suggests. Keep the defaults, add your own, or skip this step entirely — you can attach capabilities any time from the worker’s page. See Capabilities & connections for how connections become usable tools.
4

Save and start a task

Saving creates the worker and takes you to its page. Click New task (or “Chat with” the worker) to start a private task in its context. This copies the worker’s files into the new task and applies its instructions, skills, and connections automatically. Read more about how tasks work in Tasks.
5

State your goal in chat

A task starts as a conversation, not a running job. Describe what you want in plain language. For anything non-trivial, the worker may ask a clarifying question or two — and for anything with attached files, you can look at them together — before it does anything autonomous.
6

Watch it work

Once the goal is clear, the worker gets to work: it reads and writes files, runs code in its own sandbox, and calls any connections it needs. You see this happen live — plain-language step lines appear inline in the chat (“Pulled 412 invoices”, “Matched 14/14 to the ledger”), and files show up in the side panel as they’re created. If it needs to take an outbound or otherwise risky action, an approval card appears inline so you can approve, edit, or reject before it proceeds.
7

Review the output and files

When the worker finishes, you get a recap, not a wall of transcript: a one-line outcome, what changed, and any supporting evidence. From there you can approve the result, ask for changes (which drops you back into chat), or just keep the conversation going. Every file the task touched stays in its Files panel — private to you, and ready to open, download, or hand to another task.
Nothing here is one-shot. You can keep chatting with a finished task to iterate, and a worker you create today is reusable for every future task in its role.

​
Skip the worker: start a task without one

You don’t have to create a worker to get started. The top-level New task button (on the Tasks home, not under a specific worker) always opens an empty, worker-less task — the closest thing to opening a fresh chat. A worker-less task:
  • Starts you on a sensible default model, with a picker in the chat composer if you want to switch to something else — there’s no worker to inherit a model from, so you pick directly instead of filling out an upfront form.
  • Has no worker instructions, skills, or connections applied, since there’s no worker attached.
  • Still gets the core tools: it can edit files and run code in its own sandbox, search the web, and connect a capability for that task if it needs one.
  • Is private to you, exactly like a worker-bound task.
Reach for a worker-less task for a one-off request you won’t repeat. Create a worker when you expect to hand off this kind of work again — it keeps its instructions, skills, and connections ready for next time, instead of you re-explaining them every session.

​
Next steps

Anatomy of a worker

See everything that makes up a worker, and how to manage one after you’ve created it.

Capabilities & connections

Learn how integrations become tools a worker can call.

Tasks

Understand worker-bound vs. worker-less tasks and how the task lifecycle works.

Core concepts

The full Cowork vocabulary and mental model in one place.